[ docs · concepts ]

The Jev gate

The Jev gate is Grist's confidence gate. Before any model is called, it reads the task once and decides two things: which rung of the ladder should do the work, and how much effort the task deserves.

Why a gate

Most agent tools run every task on one model — usually the most expensive one, because that's the safe default. But most coding tasks are routine: a rename, a test fix, a small feature in a familiar file. Paying frontier prices for routine work is waste; paying discount prices for genuinely hard work is failure. The gate exists to tell the difference before the money is spent.

What it decides

• The rung. The cheapest rung whose capability band covers the task. See the ladder.
• The effort dial. A tight budget for a trivial, well-specified task; room to explore for an ambiguous one. Effort is priced, not improvised.
• Delegation ceilings. When a specialist subagent is spawned, its work re-gates on the delegation text — with the parent's rung as a ceiling. Decomposed work can only shed cost, never gain it.

Rules it never breaks

• Underspecified means cheap, not stuck. If a task can't be judged confidently, the gate routes it to the cheapest rung rather than escalating out of caution. Escalation is earned by evidence during the run, not assumed before it.
• It always answers. Every run gets a rung. The gate never waits on a person to make the routing call.
• It prices before it spends. The decision happens once, up front — not discovered on the invoice.

A worked example

Calibration

The gate's thresholds aren't hand-tuned folklore. Grist joins routing decisions to verified outcomes and proposes bounded adjustments, applied as configuration — the learning loop described in Dreaming & calibration. And your dashboard can override any rung's model at any time; the gate chooses among your ladder, not ours.